toadoum/ai-research-skill

AI Research Skill is a set of files you drop into your coding agent's skills directory to give it structured guidance for every stage of a machine learning research project. It works with Claude Code, Codex, and OpenClaw by placing a SKILL.md file where those agents automatically look for instructions, so the guidance loads whenever your task looks like research without any extra setup. The skill walks the agent through seven research stages: turning a vague idea into a testable hypothesis, doing a literature review from papers actually read rather than recalled, reproducing the strongest known baseline before claiming any improvement, designing experiments with proper seed fixing and data-leakage checks, running configurations so every result is reproducible, analyzing results across multiple seeds rather than a single lucky run, and writing a paper where every claim is backed by a number. What makes this project different from a static prompt is a self-improving loop. When the agent catches or makes a research mistake, it runs a small Python script that appends the lesson to a LESSONS.md file in your project. At the start of every future task, the agent reads that file and applies what it already learned. If the same mistake appears three times, the lesson gets promoted into the SKILL.md itself. The loop cannot be used to make the agent less rigorous: a built-in guardrail refuses any lesson that would weaken research integrity by fabricating, hiding, or cherry-picking results. The repository also includes reference documents for literature review, experiment design, and paper writing, plus a script that scaffolds a reproducible experiment directory with the correct folder structure. The project is licensed under MIT and aimed at researchers who run ML experiments with AI coding agents and want the agent to catch common mistakes (data leakage, single-seed results, fabricated citations) before they cost weeks of work.